Trend Observatory: Observation to Hypothesis to Pattern

1. The real problem: do not call one spike a trend

A dashboard shows a seven-day spike for a topic and two smaller increases in earlier months. The tempting move is to label it a durable trend and change the content or product plan immediately. Observation, explanation, and pattern are different layers. This lesson builds a Trend Observatory that prevents an attractive chart from becoming a decision by itself.

2. Lesson boundary

The owned artifact is Trend Observatory Evidence Ladder v1. Record an observation with a bounded source/window, check context and baseline, state a hypothesis and an alternative hypothesis, then evaluate replication, corroboration, and counter-signals. Anti-newsjacking and ethical context belong to PLN-014; experiment/change isolation belongs to PLN-016. Neither adjacent outcome is absorbed here.

3. Observation is only what you actually observed

An observation must be traceable: query or topic, region, source window, normalized interest, and observation time. “People genuinely care more now” is not an observation; it is an interpretation. A better statement is: “For topic X in region R and window W, normalized interest is higher than the declared baseline.”

4. What Google Trends does and does not show

Google documents that Trends uses a sample of searches and normalizes results to the time and location of the query; displayed values are scaled relative interest rather than absolute demand [GT-FAQ]. A value of 100 is therefore not “100 searches” and not a direct market-size measure. It is an observation about relative search interest, not proof of intent or a durable pattern.

5. Fix the baseline and comparison window before telling a story

A spike without a baseline is merely a salient point. Compare it with an appropriate preceding period, comparable historical period, or declared benchmark. Google Trends Quick Comparisons supports comparisons with preceding periods and historical references [GT-COMPARE]. If apparent growth disappears when the window changes, confidence should decrease rather than the inconvenient window being discarded.

6. Region, query/topic, and sampling note are evidence

A signal can be strong in one region and absent in another. A search term and a broader topic are not identical constructs. Keep region, query/topic choice, sampling note, retrieval time, and provenance beside the observation. A changed context creates a new observation; it must not be silently merged with the original.

7. Move from observation to hypothesis, not directly to pattern

A hypothesis is a testable explanation for the observation—for example, “the increase reflects a persistent educational need.” Record an alternative at the same time: “the increase is driven by a short-lived news event or seasonality.” If only the preferred explanation is stored, the observatory becomes a confirmation machine.

8. Replication means the signal appears again

Promotion beyond ONE_OFF requires evidence in more than one suitable observation or window. One enormous spike does not substitute for repeated compatible observations. This lesson does not impose a universal replication count; the rule is context-specific. The synthetic fixture uses explicit repeats so ONE_OFF and EMERGING_PATTERN can be tested deterministically.

9. Keep corroboration and counter-signals together

A corroborating signal is an independent or complementary measure consistent with the hypothesis. A counter-signal constrains or contradicts it. A pattern is not built by collecting only positive evidence; opposing evidence belongs in the same artifact. Dropping a counter-signal is an independent failure.

10. Seasonality and Trending Now

Google's Trending Now surface focuses on queries experiencing recent surges and exposes very recent windows through several days [GT-NOW]. That is useful for finding fresh observations, but “recent surge” is exactly why it must not be equated automatically with a stable pattern. Seasonality, event-driven spikes, and news cycles belong in alternative hypotheses or counter-evidence.

11. State machine: ONE_OFF to STABLE_PATTERN

The observatory uses five main states: ONE_OFF for a single observation, HYPOTHESIS when an explanation exists but repeated evidence is insufficient, EMERGING_PATTERN when replication/corroboration exists but important limitations remain, STABLE_PATTERN when the declared repetition/context rule is satisfied, and REJECTED when opposing evidence or a context change collapses the explanation. UNKNOWN, INSUFFICIENT_DATA, and STALE_EVIDENCE remain valid when evidence cannot support promotion.

12. Decision gate: decision strength must not exceed evidence strength

STABLE_PATTERN is not a causal truth claim. Even a stable pattern only supports bounded planning use, such as retaining a topic for the next planning horizon. Incomplete provenance, stale sources, undefined region, or unresolved counter-signals route to refresh, review, or defer. A clean dashboard does not justify removing uncertainty.

13. Worked example: an educational topic

Assume three non-overlapping windows show compatible increases, region is fixed, the baseline is appropriate, one independent signal corroborates the movement, seasonality was checked, and no material counter-signal remains. The fixture may reach STABLE_PATTERN, but the claim stays bounded: “Current evidence supports a repeated interest pattern in this context.” It does not prove user need, revenue impact, or causality.

14. Negative cases

Four shortcuts are rejected deliberately: (1) one extreme spike without replication → ONE_OFF; (2) changing region makes the signal disappear → REJECTED; (3) a seasonal explanation appears but the hypothesis is not updated → HYPOTHESIS plus review; (4) corroboration is removed while the state is kept STABLE → downgrade. Stale evidence or missing provenance also blocks promotion.

15. Transfer: growth disappears under a different window

Re-evaluate the same topic with a changed comparison window. If apparent growth disappears, do not copy the earlier conclusion. Recompute the baseline, normalized signal, alternative hypothesis, and pattern state. The contract remains fixed while context-specific evidence changes. That is what separates a trustworthy observatory from a storytelling dashboard.